Theses & student projects

Habitat imaging analysis was performed in a cohort of 38 skull-base cancer patients treated with proton therapy, of whom 7 experienced tumour recurrence. Baseline multi-modal imaging (CT, T1, FLAIR, DWI/ADC, GADO-T1 MRI) was collected. The soft-tissue component of gross tumour volumes was analysed using six clustering algorithms yielding 16 habitats. Although individual habitat volume fractions did not reach statistical significance in corrected univariate tests, multivariate survival models demonstrated predictive performance for local failures. K-Means and Otsu clustering yielded the highest concordance indices. The habitat with high GADO-T1 intensities acted as a robust hazard-increasing predictor, whereas habitats with hyperintense FLAIR or combined high ADC and FLAIR consistently demonstrated hazard-decreasing effects.

ETH supervisor: A.J. Lomax

Background and Motivation: Accurate prediction of radiation therapy response remains difficult due to inter-patient variability and tumor heterogeneity. A pipeline was developed for multimodal image registration, intensity normalization, and radiomics feature extraction from baseline DWI, T1-weighted and FLAIR MRI in 31 skull base chordoma and chondrosarcoma patients treated with proton therapy. A Cox proportional hazards model was fitted on 50 radiomics features and one clinical parameter. The final model incorporating five features achieved good risk stratification, as confirmed by Kaplan–Meier analysis. The presented pipeline provides a reproducible framework to extract radiomics features from multimodal MRI data and perform survival analysis in combination with clinical parameters.

ETH supervisor: A.J. Lomax

Correlations between local treatment failure and radiomics features derived from pre-treatment apparent diffusion coefficient (ADC) distributions within the gross tumour volume were investigated in 11 brain tumour patients, 3 of whom exhibited local treatment failure. For each patient, 58 radiomics features were extracted from normalised ADC maps registered to planning CT images. Of the considered features, only sphericity correlated significantly with local treatment failure (ρ = 0.65, p = 0.049). A number of features related to intensity heterogeneity showed moderate correlations, without reaching the significance threshold.

ETH supervisor: A.J. Lomax

Ocular proton therapy demands sub-millimeter precision in eye positioning. This study explores a refraction-aware, patient-specific modelling approach to eye pose estimation using a 3D anatomical model derived from MRI and refined via a ray-traced simulation framework. Two strategies were compared against X-ray-based ground truth. Refinement of corneal geometry and pupil centre location significantly improved alignment between simulated and observed pupil contours. Eye pose estimation using the updated model demonstrated improved accuracy particularly in vertical and beam-axis alignment, with statistically significant reductions in clip and tumour centroid deviations.

ETH supervisor: A.J. Lomax

A pipeline was developed to account for rotational positioning uncertainties in proton therapy and compare their magnitudes to translational ones, applied to a set of head and neck cancer patients treated at the Center for Proton Therapy at PSI. Dose discrepancies were evaluated for both translational and rotational uncertainty scenarios. Results showed that proton treatments of this patient site are more sensitive to translations than to rotational uncertainties, providing a way forward for improving plan robustness to ensure optimal treatment delivery.

ETH supervisor: A.J. Lomax

A mechanistic tumour control probability (TCP) model was implemented to assess the TCP increase in hypoxia-guided dose escalation proton therapy plans based on data from 10 non-small cell lung cancer patients. The model accounts for the enhanced radioresistance of hypoxic cells as well as inter-fraction reoxygenation and repopulation. Patients with the highest-degree hypoxia benefited most from dose escalation, with an increase in mean TCP of up to almost 20%. The predicted TCP for conventional dose distributions was found to be very sensitive to the efficiency of reoxygenation between fractions, a sensitivity that was decreased for dose-escalated proton plans.

ETH supervisor: A.J. Lomax

The impact of variable proton linear energy transfer (LET) on oxygen enhancement ratio (OER) estimation was evaluated in hypoxia-guided treatment planning for 10 non-small cell lung cancer patients. Variable LET was found to be 2.5 keV/µm on average in the target, versus the assumed constant value of 2 keV/µm. A 2–3 Gy difference in biologically effective dose (BED) was observed within the hypoxic subvolume between constant and variable LET simulations, while no differences were found in normoxic regions. The assumption of constant LET for OER estimation is valid in normoxic tissue but should be reconsidered for hypoxia-guided treatment planning in NSCLC.

ETH supervisor: A.J. Lomax

This study reports first steps in the development of a target decoupling system and vessel-like structures within the lung compartment of an existing lung phantom. A set-up inducing motion of a motor as a function of internal pressure was built. The latency between phantom motion and motor motion was determined to 77.67 ± 17.52 ms using an optical tracking system. A 3D printed rubber-like lung structure was examined using 4D CT and image registration. Tracking analysis of artificial vessels showed −3.3 ± 0.4 mm motion in anterior-posterior direction and 1.5 ± 0.4 mm in superior-inferior direction, suggesting more realistic motion fields compared to featureless lung phantoms.

ETH supervisor: A.J. Lomax

Polynomial, wavelet-based and support vector regression methods were investigated for respiratory signal prediction. Geometrical and temporal accuracy of predicted motion trajectories were evaluated using RMSE, jitter and phase shift metrics. Second order polynomial prediction outperforms the other algorithms with a maximal RMSE of ~0.09 mm and ~1.1 relative jitter. Polynomial algorithms were further applied for on-line detection of singular points characteristic of motion signals (full inhale peaks and end exhale plateaux). Polynomial and wavelet approaches were implemented in C, resulting in respectively ~3 µs and ~8 µs execution time.

ETH supervisor: A.J. Lomax